AI agents flagged two magnet candidates. One may have been sitting in a 1999 paper.
Vals AI says Claude Opus 5.5 agents screened materials on computers for three days and proposed two candidates for spin-based memory. One is a new design that may be hard to make. The other was first made in 1999.
Vals AI published a blog post on October 4 saying a team of Claude Opus 5.5 agents found two materials that could be useful for next-generation computer memory. The Economic Times, which covered it later, says the project used more than 90 agents over three days. The Vals post itself only says “a team of AI agents,” so I’m treating the 90 as the reporter’s figure.
What they were looking for
Here’s the goal. Ordinary magnets fall into two camps. A fridge magnet is ferromagnetic: its atomic spins line up and leak a strong field. An antiferromagnet has neighboring spins pointing in opposite directions, which cancels the field. That makes antiferromagnets dense and fast to switch, but it also makes it hard to sort electrons by spin, which is what memory needs.
A Luttinger-compensated material tries to have both: zero net magnetism and spin sorting. Vals says the sorted “spin window” has to be much wider than room-temperature thermal noise, which is about 26 meV. Wider windows mean electrons stay sorted even when the chip warms up. That’s the whole game.
Candidate one: a new design that may be hard to make
Vals says the agents designed YBaMnFeO₅, a five-element oxide that nobody had proposed as this kind of magnet. Its simulated band gap is 2.35 eV, with spin-sorted windows of 1.0 eV for holes and 1.4 eV for electrons. The simulated magnetic order holds up to about 420 K, or about 490 K after calibrating against a known magnet.
The catch shows up in Vals’ own follow-up.
The design needs manganese and iron atoms to sit in a perfect checkerboard. In the simulations, that order breaks into a random mix at about 950 K. Making this kind of oxide takes 900 to 1,300 °C, and at lower temperatures the atoms barely move, so ordinary synthesis would probably give a scrambled crystal. A scrambled crystal loses the spin sorting, which is the whole reason the material was interesting.
Candidate two: a 1999 compound nobody linked to this
KV[Cr(CN)₆] was first made in 1999. Its near-zero magnetism was built in by the chemists who made it, who designed the two metals’ moments to cancel. Vals says a 2008 study that used a similar calculation method already plotted its electron states, and both band edges carried the same spin. That study was about magnetic coupling under pressure, though, and never pointed this out. The 1999 sample stayed magnetically ordered up to 376 K, a number the original chemists measured.
The caveats matter here. The predictions assume a perfect, dry crystal. The 1999 sample was a powder with water in its pores, and it showed a small leftover magnetic moment. Vals’ two calculation methods disagree on how much the water matters. The more accurate one says spin sorting survives. The faster one says the hole window shrinks by more than half. Neither the band gap nor the spin sorting has been measured.
Why the headline is more than the data
The headline version is “AI finds a room-temperature magnet.” Honestly? I’d rewrite it. The version the data supports is “AI found two simulation candidates. One may already exist in a 1999 paper, and the next step is a lab.” Both are interesting. Only the second one is what happened.
The company running the test is also the one reporting the result. Vals deserves credit for posting its input files, raw outputs and code on GitHub, and for listing caveats in its own post. That’s more than most AI-for-science claims give you. Still, I haven’t seen anyone outside Vals rerun the calculations, and I can’t check them myself. Nobody outside the company has shown that the agents did more than a well-run search of existing work and some standard simulations, which is the honest reading of what’s public today.
My take is that this is a better demo of AI-assisted search than of a new material. Searching old papers and tables is exactly the boring work an agent can speed up. The “three days” figure measures compute, not chemistry. I’d wait for a synthesis team before calling it a memory breakthrough.
The next step Vals names is to make KV[Cr(CN)₆] again and measure its spin sorting directly. If a real sample holds up, that will matter more than any simulation. If it doesn’t, the 1999 paper goes back to being a footnote, and the agents will have found a good list of things to test rather than a new memory material. For a different kind of AI-research claim, where checking took longer than generating, see OpenAI’s 722 math papers and what Lean can and can’t verify.